cfDNA screening performance: accounting for and reducing test failures

Over several decades, antenatal screening for Down syndrome evolved steadily, from the initial concept of a single second-trimester maternal serum marker, to complex protocols involving the combination of multiple serum and fetal ultrasound markers, determined within-trimester or sequentially across trimesters. Early in this process, it was realized that the optimal way to combine markers and prior risk factors was to estimate the risk of an individual pregnancy being affected. The distributions of risk in affected and unaffected pregnancies overlap substantially, requiring a cut-off to designate results as ‘screen-positive’, requiring referral for counseling on invasive prenatal diagnosis, and ‘screen-negative’, generally with no further action. Performance was best assessed by statistical modeling to determine detection and false-positive rates. This multimarker approach is about to change radically with the discovery that a single marker, maternal plasma cell-free (cf)DNA, has a vastly superior performance to that of any of the current markers, even when combined. There have been 13 studies published to date which tested plasma samples drawn prior to invasive prenatal diagnosis (‘high-risk’ studies), seven which used samples drawn as part of existing screening programs and one with a mixture of both study designs1. Meta-analysis of the high-risk studies yielded a detection rate of 99% and a false-positive rate of 0.1%1. Performance was even better for the other studies, but these were less reliable, due mainly to incomplete follow-up, and it made no material difference whether results were based on counting DNA fragments, shotgun or targeted sequencing, or analysis of single nucleotide polymorphisms (SNPs). In contrast to cfDNA, for conventional screening, modeling predicts a detection rate of, at best, 87% for the widely used first-trimester combined test and only 71% for the second-trimester quad test, at a fixed false-positive rate of 5%2. Whilst other, less widely used, conventional screening protocols yield improved detection, this comes nowhere near 99%. Screening sequentially with the step-wise, integrated and contingent tests has a predicted detection rate of 92–94%, and extending the combined test to include further serum or ultrasound markers has a rate of 89–95%2. When carried out routinely, the predictive value (chance of being affected given a screen-positive result) for the cfDNA test is about 50%, whilst for even the best conventional protocols it is only about 2%1. However, although it can be concluded with confidence that the performance of cfDNA testing outstrips that of conventional screening, it has a number of technical and practical limitations which complicate simple comparison. These arise partly from an assumption early in test development that cfDNA would be a diagnostic test rather than just a very good screening marker. This is reflected in the very terminology used at the time, which is, to some extent, still in use today. Initially, the test was called non-invasive prenatal diagnosis (NIPD), suggesting superiority to invasive prenatal diagnosis. With the realization that there are many false results, both false positives and false negatives, the name was softened to NIPT (testing) or NIPS (screening), but part of the diagnostic claim adhered. This impression was reinforced by use of the term ‘discordant’ to describe false-positive and false-negative results emerging at invasive prenatal diagnosis or at birth. One consequence of viewing the cfDNA test as diagnostic, or almost so, was that the initial studies, as do most commercial providers today, excluded a few percent of results as ‘uninterpretable’ or ‘no-call’. One study inexplicably excluded results close to the cut-off3, whilst others had more scientific explanations of test failure. In about half the samples, the reason given was that there was too little fetal cfDNA, usually expressed in terms of the proportion of cfDNA arising from the fetoplacental unit, or the fetal fraction (FF)4. The logic behind this is the relationship between FF and the expected number of chromosome 21 copies per genome in an admixture of DNA from the euploid mother and the trisomy-21 fetus. The relationship is expressed by the formula 3FF + 2(1 – FF), or 2 + FF, per genome. So, for example, if the FF is 10%, around the average seen in early pregnancy screening, the test has to be able to distinguish 2.10 for an affected fetus from 2.00 for a euploid fetus. If the FF is very low, for example 2%, it has to distinguish 2.02 from 2.00, which may be beyond the test's capability if the aim is diagnosis, but could still provide some discriminatory value if the aim is screening. A complication of using FF to exclude samples is that this parameter is estimated with varying degrees of analytical imprecision. Furthermore, the smaller the actual FF, the more imprecise will be the estimated value. This means that a relatively high minimum acceptable value for measured FF may be required to avoid accepting a very low actual FF. Most providers use a minimum of 4%. There are now considerable data available on the practical aspects of cfDNA screening by commercial providers, that report sample exclusion rates of 1.6% (Verinata5 and LifeCodexx (Michael Lutz, pers. comm.)), 1.9% (Sequenom6), 2.2% (BGI7 and Ariosa (Thomas Musci, pers. comm.)) and 6.4% (Natera8). Given that the chance of exclusion is similar in Down syndrome and unaffected pregnancies, the meta-analysis detection rate of 99% is equivalent to effective detection of only 93–97% when the whole population is considered, including the failures. There are a number of options available for women with cfDNA failure, which will increase the effective detection rate but at the cost of a large increase in the false-positive rate, including: invasive prenatal diagnosis, detailed ultrasound examination, repeat testing and conventional screening. Although the cfDNA test failure rate is not increased in Down syndrome, among test failures there is an increased chance of another aneuploidy, particularly Edwards syndrome or triploidy, presumably due to the associated small placentae and hence lower circulating levels of fetal cfDNA (see references cited by Benn et al.9). Some women regard the chance as being sufficiently high as to warrant invasive prenatal diagnosis, but the results of a detailed ultrasound examination may be sufficiently reassuring to avoid this in most cases. If the reason for cfDNA test failure is low FF, there is a high chance of a repeat failure unless the sample redraw interval is extended to allow sufficient time for the placenta to enlarge and the FF to increase substantially. This is borne out by results from three large series in which the repeat failure rates were 27%6, 36%8 and 37%10; the providers of these tests had an average test result turnaround time of about a week. Conventional screening is another option for those with failed cfDNA tests. A locality replacing the combined test by routine cfDNA testing is likely to retain ultrasound nuchal translucency (NT) determination to detect other congenital abnormalities11. However, modeling predicts that the Down syndrome detection rate using NT alone is, at best, 77% for a 5% false-positive rate2. Detection can be increased in these circumstances using a ‘reflexive’ combined test: two blood samples are initially drawn for all women: the first is tested for cfDNA and the second analyzed for serum markers if the cfDNA test fails12. There are a number of factors associated independently with FF, including positive correlation with gestational age and combined test serum marker levels, and negative correlation with maternal weight, maternal age, South Asian ethnicity and assisted conception10. The positive correlation between FF and gestational age is likely to be due to an increase in levels of fetal cfDNA with the growing placenta; the negative correlation with maternal weight could result from the larger amounts of maternal cfDNA; and assisted conception could result in low FF by causing impaired placentation. In addition, a recent case report described a woman with a Down syndrome pregnancy having a cfDNA test at 19 weeks' gestation whilst on low-molecular-weight-heparin therapy and at 21 weeks following 24 hours of heparin cessation13; the FF increased between the timepoints from 1.6% to 5.8%, suggesting that heparin reduces the amount of fetal cfDNA, perhaps due to there being less trophoblast apoptosis. Two factors, gestational age and maternal weight, could be taken into account in order to reduce the failure rate of cfDNA testing. Restricting testing to gestational ages of 10 weeks or more would reduce the number of cases with low FF14. A study of maternal weight and gestation concluded that in Class-II or -III obesity, cfDNA testing should be delayed until 21 weeks, although it was estimated that the failure rate for this group would still be 10% compared with 4.4% across all weights and gestations15. The failure rate of the commercial SNP-based cfDNA test is much higher than that of the methods based on counting DNA fragments, but is reduced when an optional paternal buccal sample is also taken8, 14. This suggests a policy of always taking a paternal sample when using this method in obese women. A study examining this approach found that the paternal sample reduced the failure rate in obese women from 33% to 12%, but this remained much higher than the reduced rate of 1.8% (from 5.0%) in non-obese women16. Another option, for all cfDNA methods, would be to anticipate a high failure rate in obese women and ensure that a serum sample is also taken for reflexive testing. Conventional multimarker screening protocols have lower detection rates in twin pregnancies discordant for Down syndrome than they do in singletons. This is due to an abnormal serum concentration from the affected fetus being masked by that from the unaffected cotwin. A similar problem arises with cfDNA screening. Assuming the FF to be the same for each fetus, the expected number of chromosome 21 copies per genome is (3 × 0.5FF + 2 × 0.5FF) + 2 × (1 − FF) or 2 + 0.5FF, compared with 2 + FF in singletons. Although offset to some extent by the higher FF in twins, estimated to be one-third in one study17, detection will necessarily be lower. In practice, the FF will differ between the fetuses and, if the affected cotwin has the smaller value, detection will be even lower. To avoid this, a policy has been adopted whereby both FFs are estimated and the smaller is compared with the minimum acceptable value. This means that the failure rate will be higher in twins than in singletons: 9.4% and 2.9% in one study18. The widely adopted minimum acceptable measured FF value of 4% may not be applicable to all cfDNA methods, and it has been suggested that each provider should choose a value based on their own limit-of-detection19. To illustrate the point, the authors mixed samples from trisomy-21 fetuses with samples from non-pregnant euploid women and tested for cfDNA, using a targeted counting method. The lowest FF yielding a positive cfDNA result was found to be 2% and this was regarded as the limit-of-detection. Moving the minimum acceptable value to this limit would reduce the number of failures due to low FF but there is no guarantee that the detection rate would not also be reduced; a large prospective study would be needed to quantify any reduction. Improvements in analytical precision will also reduce the failure rate of cfDNA testing. For the counting methods, one approach is to increase the ‘depth of sequencing’, i.e. the number of DNA fragments aligned to a given chromosome. A theoretical model was used to assess the consequences of increasing the depth of sequencing according to FF20. At a depth of one million DNA fragments, the predicted Down syndrome detection rate at FFs of 3%, 5%, 10% and 20% were 14%, 57%, 99% and 100%, respectively. Increasing the depth to five million, the predicted detection rate was 100%, except at a FF of 3%, in which case it was 89%; with an increase in the depth to 10 million, even at a FF of 3% the predicted detection rate was 100%. Another improvement to counting methods is ‘paired-end’ sequencing, which allows fragment length to be determined and separate counting of small fragments, which are more likely to be from fetal cfDNA21. Other technical advances likely to improve precision are currently in progress. One commercial company using a standard counting method has developed a fully automated test which yields a low failure rate22. One company that initially used sequencing has replaced this with a microarray, thereby eliminating an expensive step23. Another company has presented proof-of-principle results using an automated method that eliminates the expensive steps of sequencing, polymerase chain reaction (PCR) and complex data analysis: the cfDNA fragments are captured, converted into DNA circles, amplified and labeled with chromosome-specific flourophores, and counted on a microtiter plate24. Removal of PCR from the process will improve precision, and counting unique molecules will lead to a much higher depth of sequencing than that achieved with current cfDNA methods. Another approach analyzes differentially methylated markers on chromosome 21 by immune-precipitation and real-time PCR. The discriminatory power of the test depends on the enrichment ratio of each marker in an affected pregnancy compared with euploid controls. These ratios have been shown to be independent of FF, which implies that FF need not be determined in order to interpret the test result25. The question arises whether there could be sufficient improvement in precision as to obviate the need for FF determination, or whether this will always be needed to preclude the extremely unlikely event of a sample mix-up with one from a non-pregnant woman26. Would it be sufficient for any improved, more precise, method to demonstrate that, in women with FF below 4% (say), the Down syndrome detection rate for cfDNA exceeded that of the combined test? Or, more practically, would it be sufficient for the method to demonstrate an overall detection rate of 99% without excluding low FF samples, with 1% failure for other reasons? This might be preferable to a current method with a detection rate of 99%, but 2% failures, half being due to low FF. The effective detection rate for the new, more precise, test would be 98% compared with 97% for the current test, and resolution of failures through repeat testing and conventional screening would likely be more complete for the precise test. So far, the high cost of the cfDNA test has precluded its routine introduction in public health settings. Instead, two cost-effective options are being considered (see review by Ginsberg and Cuckle27): (1) a secondary cfDNA test offered to those with a screen-positive conventional screening result; and (2) a contingent cfDNA test for the 15–25% of women with the highest Down syndrome risk based on conventional screening. Secondary screening is cost-neutral or better, but will result in a reduction in the conventional test detection rate assuming that compliance of invasive prenatal diagnosis is unaltered. Modeling predicts that contingent screening with 20% of women selected for cfDNA by a combined test would yield a detection rate of 94%1. With both options, the consequences of cfDNA test failure are ameliorated by the presence of the conventional test result. It is expected that, in the near future, competition between the existing providers and newer methods will reduce costs and open up the prospect of routine cfDNA testing. At that point, public health decision-makers will have to take into account the practical issue of managing cfDNA test failures. It will be critical to minimize this problem whilst maximizing the effective Down syndrome detection rate. H.C. is a consultant to PerkinElmer Inc, a producer of reagents for conventional screening analytes which is developing a non-sequencing cfDNA method, and a director of This is My Ltd, a provider of conventional and cfDNA screening services.